squashed commit

.gitignore

@@ -87,6 +87,9 @@ target/
 profile_default/
 ipython_config.py
 
+# generated files
+**/generated/**
+
 # pyenv
 #   For a library or package, you might want to ignore these files since the code is
 #   intended to run in multiple environments; otherwise, check them in:

CMakeLists.txt

@@ -224,6 +224,48 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
           "-gencode arch=compute_90a,code=sm_90a")
   endif()
 
+  #
+  # For the CUTLASS-Marlin (marlinv2, mixed precision cutlass kernels) we 
+  # automatically generate sources for various preselected input type pairs
+  # and schedules.
+  # Generate sources:
+  execute_process(
+    COMMAND ${CMAKE_COMMAND} -E env 
+    PYTHONPATH=${CMAKE_CURRENT_SOURCE_DIR}/csrc/cutlass_extensions/:${CUTLASS_DIR}/python/:$PYTHONPATH 
+      ${Python_EXECUTABLE} ${CMAKE_CURRENT_SOURCE_DIR}/csrc/quantization/marlinv2/generate.py
+    RESULT_VARIABLE marlinv2_generation_result
+    OUTPUT_VARIABLE marlinv2_generation_output
+    OUTPUT_FILE ${CMAKE_CURRENT_BINARY_DIR}/marlinv2_generation.log
+    ERROR_FILE ${CMAKE_CURRENT_BINARY_DIR}/marlinv2_generation.log
+  )
+
+  if (NOT marlinv2_generation_result EQUAL 0)
+    file(READ ${CMAKE_CURRENT_BINARY_DIR}/marlinv2_generation.log log)
+    message(FATAL_ERROR "MarlinV2 generation failed."
+                        " Result: \"${marlinv2_generation_result}\"" 
+                        "\nCheck the log for details: "
+                        "${CMAKE_CURRENT_BINARY_DIR}/marlinv2_generation.log")
+  else()
+    message(STATUS "MarlinV2 generation completed successfully.")
+  endif()
+
+  # Add marlinv2 generated sources
+  file(GLOB MARLINV2_GEN_SOURCES "csrc/quantization/marlinv2/generated/*.cu")
+  list(APPEND VLLM_EXT_SRC ${MARLINV2_GEN_SOURCES})
+  message(STATUS "MarlinV2 generated sources: ${MARLINV2_GEN_SOURCES}")
+
+  # See comment above for scaled_mm_c3x (same if condition)
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0)
+    set_source_files_properties(
+          ${MARLINV2_GEN_SOURCES}
+          PROPERTIES
+          COMPILE_FLAGS
+          "-gencode arch=compute_90a,code=sm_90a")
+  endif()
+
+  # Add pytorch binding
+  list(APPEND VLLM_EXT_SRC
+    csrc/quantization/marlinv2/marlinv2_pytorch.cu)
 endif()
 
 define_gpu_extension_target(